`nuw` being lost when optimizing `sub` to `add` leads to worse codegen

[JakobDegen]

See the problem on godbolt. In this code:

define i64 @fast(i8* align 1) unnamed_addr #0 {
  %2 = load i8, i8* %0, align 1
  %3 = sub nuw i8 %2, 10
  %4 = zext i8 %3 to i64
  ret i64 %4
}

running through opt causes the sub to be switched out for

%3 = add i8 %2, -10

Note that the nuw flag was lost. This causes the compiled assembly to include an extra instruction. Originally found in rust-lang/rust#91449 .

Alive2 actually indicates that the nuw flag would be incorrect here, which makes sense since -10 is secretly just 246. Is there another way to address this? I can close this issue if the only reasonable response is "yeah, canonicalization requires us to do this, and sometimes that causes problems. Sorry"

[nikic]

Yeah, this is a known problem. I've hit this issue many times.

The only generic way to address this that I see is to add a new snuw (or so) flag that basically say that it's nuw after converting add x, y into sub x, -y.